hermes - 💡(How to fix) Fix memory: No archive path when removing entries from bounded MEMORY.md — deletion is permanent, no bridge to Hindsight [1 pull requests]

Problem

When MEMORY.md reaches its character limit (default 2,200, configurable), the memory tool blocks new add operations, forcing users to manually delete old entries via memory(action="remove"). These deleted entries are permanently lost — there is no archive, offload, or bridge path to any long-term memory backend (Hindsight, Honcho, etc.).

The Categorical Gap

The built-in memory tool and Hindsight serve fundamentally different purposes:

They are orthogonal, not redundant. Hindsight cannot replace MEMORY.md, and MEMORY.md deletions cannot be recovered from Hindsight.

Evidence from a production deployment

• MEMORY.md: 3,490/3,500 chars (99.7% full), 42 entries
• Hindsight: 28,713 records (daemon PID 1601480, healthy)
• 4 entries have NO redundancy elsewhere (not in SOUL.md, USER.md, or AGENTS.md):
  1. Feishu MEDIA path whitelist + format restrictions + kanban.db root cause
  2. Network proxy configuration (Codex→8118, Stotik→7890, PubMed→直连)
  3. Cherry-pick iron rule (PR→cherry-pick→push→switch→delete→verify→restart)
  4. AI fabrication prevention (Grok/DeepSeek systematic fabrication, countermeasures)

Recall Verification

8 targeted hindsight_recall queries for memory rule content returned event records (cherry-pick operations, PR activity, Kanban notifications) — never the rule text itself. Hindsight stores "what the agent DID", not "what the agent MUST do".

Root Cause

1. on_memory_write() hook exists in MemoryProvider ABC (agent/memory_provider.py:274) but Hindsight does NOT implement it — only sync_turn() (conversation auto-extraction).
2. auxiliary.flush_memories was removed in v0.12.0 (PR #15696, see RELEASE_v0.12.0.md:169) — the config block in config.yaml is dead.
3. memory remove has no archive path — the entry is popped from the list and the file is persisted. No hook fires.

Steps to Reproduce

1. Fill MEMORY.md close to memory.memory_char_limit
2. Add a new entry → receives "success": false, "error": "Memory at X/Y chars. Adding this entry... would exceed the limit."
3. Run memory(action="remove", old_text="<existing entry>") → entry is deleted from file
4. Query Hindsight for the deleted entry's content → returns event records, NOT the rule text
5. The deleted entry is irrecoverable

Suggested Fix

Implement on_memory_write() in HindsightMemoryProvider to auto-archive removed entries:

def on_memory_write(self, action, target, content, metadata=None):
    if action == "remove":
        self._retain_queue.put({
            "text": f"[ARCHIVED from {target}] {content}",
            "tags": ["archived", f"source:{target}"],
        })

This preserves the entry in Hindsight with proper tags, making it retrievable via hindsight_recall even after removal from MEMORY.md.

Alternative approaches:

• Add action="archive" to the memory tool: read entry → hindsight_retain → remove
• Warn when MEMORY.md exceeds 90% capacity (preventive)
• Surface the gap in the overflow error message: "Consider using hindsight_retain to archive entries before removing them"

Related Issues

• #28424 "Memory: auto-compact MEMORY.md when char limit reached" — addresses LLM summarization, not Hindsight offload
• #32064 "Fix bounded memory: durable unlimited user memory with retrieval" — addresses overflow write paths, not deletion archival
• #23378 "Bounded memory overflow does not trigger automatic consolidation" — addresses auto-consolidation, not archive bridge
• #15696 removed flush_memories entirely

Environment

• Hermes Agent: server-mainline branch (synced with upstream/main)
• Memory provider: hindsight (local_external mode, port 8888)
• MEMORY.md: 3,490/3,500 chars
• Hindsight: 28,713 records